Something's wrong with the new Glaurung

[Vempele]

In the starting position, with a single thread. The results are reproducible. 32-bit:

Code: Select all

Glaurung 071001.  Copyright (C) 2004-2007 Tord Romstad.
go depth 10
info depth 2
info depth 2 score cp -9 time 3 nodes 45 nps 15000 pv g1f3 b8c6
info depth 3
info depth 3 score cp 29 time 7 nodes 193 nps 27571 pv g1f3 b8c6 b1c3
info depth 4
info depth 4 score cp -9 time 12 nodes 353 nps 29416 pv g1f3 b8c6 b1c3 g8f6
info depth 5
info depth 5 score cp -11 time 17 nodes 653 nps 38411 pv g1f3 b8c6 b1c3 g8f6 h2h
4
info depth 6
info depth 6 score cp -11 time 23 nodes 1258 nps 54695 pv g1f3 b8c6 b1c3 g8f6 h2
h4 d7d6
info depth 7
info depth 7 score cp -17 time 30 nodes 3132 nps 104400 pv g1f3 b8c6 b1c3 g8f6 h
2h4 d7d6 e2e3
info depth 8
info depth 8 score cp -13 time 39 nodes 5763 nps 147769 pv g1f3 b8c6 b1c3 g8f6 d
2d3 f6g4 c3d5 h7h5
info depth 9
info depth 9 score cp -19 time 57 nodes 14500 nps 254385 pv g1f3 b8c6 b1c3 g8f6
e2e3 h7h5 f1b5 a7a6 b5c6 b7c6
info depth 10
info depth 10 score cp -11 time 99 nodes 40040 nps 404444 pv g1f3 b8c6 b1c3 g8f6
 e2e3 e7e6 h2h4 h7h5 a2a3 f6g4
bestmove g1f3 ponder b8c6

64-bit:

Code: Select all

Glaurung 071001.  Copyright (C) 2004-2007 Tord Romstad.
go depth 10
info depth 2
info depth 2 score cp 11 time 2 nodes 46 nps 23000 pv g1f3 g8f6
info depth 3
info depth 3 score cp 54 time 6 nodes 185 nps 30833 pv g1f3 g8f6 b1c3
info depth 4
info depth 4 score cp 11 time 11 nodes 369 nps 33545 pv g1f3 g8f6 b1c3 b8c6
info depth 5
info depth 5 score cp 15 time 16 nodes 771 nps 48187 pv g1f3 g8f6 b1c3 b8c6 d2d3

info depth 6
info depth 6 score cp 11 time 21 nodes 1577 nps 75095 pv g1f3 g8f6 b1c3 b8c6 d2d
3 d7d6
info depth 7
info depth 7 score cp 25 time 28 nodes 3543 nps 126535 pv g1f3 g8f6 b1c3 b8c6 e2
e3 d7d6 f1d3
info depth 8
info depth 8 score cp 11 time 36 nodes 6186 nps 171833 pv g1f3 g8f6 b1c3 b8c6 e2
e3 e7e6 f1d3 f8c5
info depth 9
info depth 9 score cp 56 time 55 nodes 20783 nps 377872 pv g1f3 g8f6 b1c3 b8c6 e
2e3 c6b4 f3e5 b4c6 e5c6 b7c6
info depth 10
info depth 10 score cp 11 time 76 nodes 35492 nps 467000 pv g1f3 g8f6 b1c3 b8c6
e2e3 e7e5 f1d3 f8c5 e1g1 e8g8
bestmove g1f3 ponder g8f6

[Tord Romstad]

Teemu,

Thanks for letting me know. This isn't a serious problem -- there is no good reason why anyone would want to use this Glaurung version. The only reason it exists at all is to test the scaling on 8 CPUs, which Daniel has now helped me do.

When it's time to release Glaurung 2.0, I'll make sure to test that it works on both 32-bit and 64-bit systems.

Tord

[Eelco de Groot]

Hi Tord I was wondering whether you can confirm that the Recapture extensions in Non-PV nodes are not yet working? I could see no difference when I increased them from zero to 1

Also I was wondering if you have ever tested a configuration with all the Pawn-push to 7th extensions, for both PV and Non-PV nodes, increased to 2 together with Single Reply extensions in PV nodes from 1 to two? It seemed not very bad but is it too early to try a few games with this version.

It is looking good so far!
Regards, eelco

[Tord Romstad]

Hi Tord I was wondering whether you can confirm that the Recapture extensions in Non-PV nodes are not yet working? I could see no difference when I increased them from zero to 1

Hello Eelco,

You are right: You've found a small bug. The bug isn't really that recapture extensions don't work at non-PV nodes, but rather that the UCI parameter "Recapture Extension (non-PV nodes)" shouldn't have been there at all. Recapture extensions at non-PV nodes hurt the playing strength so much that it is pointless to include them, even as an optional feature.

Thanks for reporting the bug!

Also I was wondering if you have ever tested a configuration with all the Pawn-push to 7th extensions, for both PV and Non-PV nodes, increased to 2 together with Single Reply extensions in PV nodes from 1 to two? It seemed not very bad but is it too early to try a few games with this version.

Extending by a full ply (which means a setting of 2) for pawn pushes to the 7th rank might work OK. I would guess that it has little or no effect on playing strength, but would solve a few test positions more quickly. Extending by a full ply for single replies ply at PV nodes looks a bit excessive to me, but I don't know.

It is looking good so far!

Thanks, but actually it isn't very exciting. It's basically just Glaurung 2-ε/5 with a more selective search, more complete UCI support, better time managment, working Chess960 support, better scaling on multiple CPUs, and a new license.

Tord